ECTI Transactions on Computer and Information Technology (ECTI-CIT)

Enhancement of Machine Learning Algorithm in Fine-grained Sentiment Analysis Using the Ensemble

2025-01-02T11:07:37+07:00

Fine-grained sentiment analysis plays a crucial role in extracting subtle opinions from textual data, especially in domains such as customer reviews and social media analysis. Traditional machine learning models, including Support Vector Machines (SVM), Naïve Bayes, and Decision Tree, often face limitations in accurately classifying fine-grained sentiments due to their inability to generalize well in complex classication tasks. To address this challenge, this study proposes an ensemble learning approach integrating voting, bagging, boosting, and stacking to enhance sentiment classification performance. Experiments were conducted on multiple datasets, comparing standalone classiers and ensemble-based approaches. The results indicate that stacking-based ensemble models achieve the highest accuracy, reaching 92.45%, outperforming traditional classiers such as SVM (88.23%) and Naïve Bayes (85.67%). Additionally, ensemble methods demonstrate improved generalization and robustness, reducing misclassification rates by 6% on average compared to individual classifiers. Among the tested ensemble techniques, stacking consistently delivered superior results, leveraging diverse weak learners to optimize sentiment classication accuracy. This research highlights the eectiveness of ensemble learning in fine-grained sentiment analysis, oering a robust methodology for improving classication accuracy and reducing sentiment misclassication. The ndings suggest that ensemble approaches, particularly stacking, provide a more reliable and scalable solution for sentiment analysis tasks, making them suitable for real-world applications in natural language processing.

Development of a Semantic Ontology for Knowledge of Ancient Lanna Documents

2025-01-02T10:53:01+07:00

This study is a research and development project aimed at developing a semantic ontology for ancient Lanna documents. It focuses on building a structured relationship of knowledge by analyzing information from various documents and databases, including the research database, the online information resource database (OPAC), the Northern Thai Information Center of the Chiang Mai University Library, the online information resource database of Rajamangalaphisek National Library, Chiang Mai, the online information resource database (OPAC) of the National Library of Thailand, the Oce of Arts and Culture, Chiang Mai Rajabhat University, and the online database from Sirindhorn Anthropological Center. This research tackles inefficiencies in retrieving and managing information on ancient Lanna documents through the development of a semantic ontology. The aim is to enhance the organization, classification, and accessibility of these documents, thereby improving search capabilities and knowledge dissemination. The innovation of this research lies in the development and evaluation of a semantic ontology specifically tailored for ancient Lanna documents. This ontology facilitates more effective grouping, categorization, and retrieval of information, significantly enhancing access to and utilization of ancient knowledge. A key innovation of this research is the application of ontology and semantic web technologies to the study of ancient Lanna documents. The research presents a structured approach to developing and validating the ontology, utilizing tools like Protégé and involving expert evaluations to ensure accuracy and relevance. The research process is divided into three stages. Stage 1 involves determining the need for an ontology by analyzing online data and grouping related keywords and phrases through the study of various information resources, including digital collections of ancient Lanna documents. Stage 2 focuses on developing the ontology using the Protégé program, which involves designing classes, setting main classes, subclasses, hierarchies, and properties to create data relations within each class. Stage 3 encompasses the ontology assessment, which is divided into two parts: evaluating the appropriateness of the ontology structure by experts through a questionnaire on class correlation validity and assessing word grouping in ancient Lanna documents. The study's findings indicate that the identication of denitions, scope, and development objectives is appropriate (mean score = 0.88), with high scores in class grouping and ordering (score = 0.90), naming relationships and properties (score = 0.90), and the overall preciseness and appropriateness of the ontology development for ancient Lanna documents (score = 0.89).

Enhanced Hand Vein Segmentation Using Generative Adversarial Network Integrated with Modied ECA Module

2025-01-16T15:14:14+07:00

Hand vein image segmentation is crucial for diverse applications such as precise biometric identification and facilitating medical intravenous procedures. This paper introduces an enhanced hand vein image segmentation method utilizing deep learning, specifically a conditional generative adversarial network (cGAN). The cGAN is trained adversarially and augmented with a modied ecient channel attention (ECA) mechanism module. The efficiency of the proposed technique was evaluated using four hand vein datasets: self-acquired dataset, SUAS, WILCHES, and BOSPHORUS. Performance comparison reveals that the proposed method outperforms alternative approaches, achieving the best results across all datasets with an average sensitivity of 0.8878, average accuracy of 0.9639, and average dice coeffcient of 0.7904 for vein patterns. Our experimental findings demonstrate that the proposed segmentation technique significantly enhances hand vein patterns and improves dorsal hand vein detection accuracy.

Machine Learning Model for Predicting the Suitability of Cultivating Alternative Crops in Lower Northern Thailand

2024-12-26T17:12:20+07:00

Intensive rice cultivation presents significant environmental and economic challenges. While crop diversification offers potential benefits for agricultural sustainability and financial resilience, farmers face considerable uncertainty when transitioning to alternative crops. This study assessed the prediction efficacy of machine learning (ML) models in identifying suitable crops for cultivation in a specific geographical area considering various factors influencing agricultural viability. Through comprehensive experimentation, a decision tree model, an artificial neural network (ANN), and a Naïve Bayes model were used for predictions and rigorously evaluated for various crops, including rubber, coconut, longan, durian, rambutan, and mangosteen. Various hyperparameter configurations were tested, and multiple evaluation indicators were employed to assess the prediction performance of the models. The results consistently demonstrated the superiority of the decision tree model, which exhibited high accuracy, precision, recall, and F-measure across most crops. Its ability to capture intricate patterns and relationships between crop attributes and suitability levels underscores its value as a decision-support tool in agriculture. While the ANN model performed well for coconut, its effectiveness varied across the other crops, highlighting the need for tailored model selection. This study provides valuable insights into the application of ML in agricultural decision-making processes, suggesting potential avenues for future optimization and enhancement of prediction accuracy.

Multi-Task Learning with Fusion: Framework for Handling Similar and Dissimilar Tasks

2025-01-16T14:58:25+07:00

Multi-Task learning (MTL), which emerged as a powerful concept in the era of machine learning deep learning, employs a shared model trained to handle multiple tasks at simultaneously. Numerous advantages of this novel approach inspire us to instigate the insights of various tasks with similar (Identification of Sentiment, Sarcasm, Hate speech, Oensive language, etc.) and dissimilar (Identification of Sentiment, Claim, Language) genres. This paper proposes two Multi-Task Learning (MTL) framework schemes based on Bidirectional LSTM (BiLSTM) to handle both similar and dissimilar tasks. The performance of these frameworks is evaluated and compared against standalone classifiers, demonstrating their effectiveness in improving classification accuracy. In order to train our proposed MTL frameworks, different task-related publicly available datasets were collected, and each sentence was annotated with all task labels with the help of publicly available pre-trained models. Along with a simple MTL framework, this paper presents an MTL framework with a fusion technique (MTL fusion) that combines learning from task-specific layers to make predictions. Our proposed MTLfusion framework provides an F1 score of 0.76, 0.92, 0.809, 0.798, and 0.89 for sentiment, sarcasm, irony, hate speech, and offensive language classification tasks, respectively (similar tasks). It also provides an F1 score of 0.59, 0.586, and 0.707 for claim, sentiment, and language identification tasks, respectively. Our research also shows that MTL frameworks perform better than their corresponding standalone classifiers for similar tasks. On the other hand, for dissimilar tasks, the standalone classifiers perform better than MTL frameworks.

Cost-Ecient Sequential Predictions with a Hybrid Method of Topological Sorting and Boosting: A Case Study on LDPE-Property Prediction

2025-03-11T11:41:56+07:00

Determining low-density polyethylene (LDPE) properties typically requires extensive laboratory testing, which is time-consuming and costly. For instance, the conditioning phase alone for measuring Vicat softening temperature requires a minimum of 40 hours [1]. Predictive modeling can reduce these costs. Ensuring accuracy that meets manufacturing standards, however, is challenging. This paper introduces TopSABoost, a hybrid method that combines topological sorting and boosting techniques to perform sequential predictions of LDPE properties and minimize the overall laboratory-testing cost. This approach reduces laboratory testing costs by predicting one property first and using it to predict another. The complexity analysis demonstrates that the proposed algorithm is ideal for non-real-time determination of sequential predictions, as it computes the model offline once for repeated use without requiring recalculations, aligning with manufacturing needs. The experimental results demonstrate that TopSABoost achieves a maximum error of just 0.11%, satisfying strict manufacturing constraints. TopSABoost identifies that prioritizing the prediction of L-value, followed by Density, offers the most cost-efficient sequence and significantly reduces reliance on direct laboratory testing while maintaining adherence to error thresholds.

Deep Learning Based Wheat Yield Prophecy and Irrigation Schedule Management to Reduce Water Waste

2025-01-16T15:54:06+07:00

Water is an incredibly valuable resource on our earth; however, it could have threatened if not managed. The agriculture has the highest necessity for strategies to minimize water usage. Agriculture industry is implementing contemporary farming methods, and farmers are using cutting-edge digital innovations that are modernize decision-making and protability in agriculture. Numerous sectors have experienced the effective use of deep learning (DL) in the decision-making. There is impetus to use it in other significant fields like agriculture. Estimating yields is essential for managing crops, water planning, ensuring food safety, and determining how much work will be needed for the cultivation and storing of crops like wheat. Predicting wheat crop yield has the potential to diminish energy use like drop in water consumption. In this study, a deep reinforcement learning (DRL) model is implemented to forecast wheat crop yield by monitoring the environment via a DRL agent. Two bidirectional long short-term memory (BiLSTM) models are applied as the DRL agent for exploring the environment. One forecasts the water content in the land and other one was active to calculate the yield considering climate data, growth stage, growing degree days (GD), canopy cover (CC), standard evapotranspiration (ETo), irrigation level and water content in soil. The agent was trained to plan watering for a wheat crop, considering a place in Maharashtra, India. DRL agent provides a schedule identifying irrigation levels. The irrigation level is incorporated into the time required to water the area, facilitating the farmer to manage it more easily. The performance of the proposed model was compared to a xed base irrigation system. Water use decreased by 35% and wheat crop output increased by 5% when the trained model was compared to the fixed technique.

Sarcasm Messages Detection using Hybrid Features Extraction Deriving from Context and Content Sentences on Social Networks

2025-03-22T12:01:41+07:00

This research aims to enhance the detection of sarcastic messages in the Thai language across social networks. It involves extracting and analyzing context-based features from messages to identify and differentiate sarcastic content. The study employs deep learning and machine learning techniques to classify these messages. The experimental findings demonstrate that a combination of context-based and content-based features yields the highest accuracy in identification. Specifically, the utilization of a bidirectional Long-Short Term Memory (Bi-LSTM) with 256 nodes, ReLU as the activation function, a dropout rate of 0.2, Sigmoid as the output activation function, binary cross-entropy as the loss function, and the Adam optimizer resulted in the highest accuracy achieved by the Bi-LSTM model, reaching 96.79%.

Mixed Reality Application for Virtual Tourism: Annah Rais Longhouse

2025-02-20T18:11:24+07:00

The tourism industry worldwide has faced unprecedented challenges recently, which has increased the need for innovative solutions to showcase destinations sustainably. Virtual tourism has emerged as a transformative response to these challenges, allowing individuals to explore destinations remotely through immersive digital experiences. Malaysia is renowned for its breathtaking natural landscapes and diverse cultural heritage, such as the Annah Rais Longhouse in Sarawak. Despite the decline in tourist numbers, there are concerns about the degradation of Sarawak's natural sites. Furthermore, individuals with disabilities often encounter barriers when trying to access these locations, which can exacerbate feelings of exclusion and isolation. This research focuses on developing a mixed reality (MR) application that leverages immersive technology to provide virtual tourism experiences, preserve natural attractions, and enhance accessibility for all types of visitors in response to these issues. The research adopted the multimedia development life cycle encompassing ve key stages: concept, design, data gathering, assembly, and testing. The developed MR applica- tion offers users two modes: Explore Mode and Virtual Tour Mode. With the integration of interactive elements, the MR application allows users to delve deeper into the history and significance of the cultural heritage, which positively contributes to the tourism industry in Sarawak. The developed MR application was evaluated using the constructs of perceived ease of use (PEOU), perceived usefulness (PU), perceived interaction (PI), and telepresence (TP). Descriptive analyses of the constructs were conducted. The results demonstrated that the MR application is well-accepted.

Machine Learning-Based Adaptive Equalization with Software-Dened Radio Experimental Setup

2025-03-29T12:04:19+07:00

High-speed data transfer over the communication channel is now possible due to developments in wireless communication. However, as these data are transmitted over the channels, multiple elements' interference and interventions will disrupt the network's functionality, frequently causing data to be misinterpreted or distorted owing to overlap. Channel equalization is a notion that can be applied with the help of machine learning and artificial intelligence to counteract this kind of interference. The hybrid technique, which extracts features from an equalizer utilized in the channel in training and tracking modes, is the subject of current research efforts. Machine learning techniques are applied to distinguish between low, high, medium, and open space situations. Analysing the radio frequency signals that travel across the channel allows for distinction. The outcome and examination of multiple machine algorithms show that the suggested model functions well in a real-time setting. Multiple sample ratios and classifier models are used to train and test algorithms such as SVM, decision tree, random forest, KNN, logistic regression, and naive Bayes. Based on the parameters mapped by the confusion matrix, machine learning algorithms' performance and efficiency are estimated. When there are fewer samples overall, the random forest algorithm performs better than other algorithms. When there are more samples, the tree-based approach produces superior results. Decision trees can be implemented in real time since, in comparison, all environment types in high, low, and medium cluttered environments have produced better results.

Energy-Efficient Per-Core DVFS for Virtual Machine Management in Cloud Data Centers

2025-01-16T15:10:15+07:00

Energy efficiency and thermal management are critical challenges in virtualized cloud data centers, particularly for optimizing parallel workloads. Dynamic Voltage and Frequency Scaling (DVFS) is widely used to balance power consumption and computational performance. This study proposes an Adaptive Threshold Per-Core DVFS Governor that dynamically adjusts CPU core frequencies based on per-core utilization, improving energy efficiency and workload performance. The proposed algorithm is evaluated using Charm++ parallel workloads and is benchmarked against existing Linux governors, including the Conservative, OnDemand, and Performance governors. Experimental results demonstrate that the proposed approach achieves superior energy efficiency per Giga-instructions compared to the Conservative and OnDemand governors while maintaining performance levels comparable to the Performance governor. Furthermore, the proposed method reduces average CPU temperature by approximately 5% (2.5◦C lower) compared to the Performance governor, contributing to enhanced thermal management in cloud computing environments. These findings highlight the potential of the adaptive per-core DVFS mechanism for improving energy efficiency and performance in virtualized data centers.

Detection of Dangerous Motorcycling Using YOLO and Machine Learning Classifiers

2025-03-29T10:22:00+07:00

The work studied three methods to identify risky motorcycle riding. It aids in identifying risky motorcyclists who adopt unusual riding positions, which lead to a rise in traffic accidents. This work established the feasibility of monitoring hazardous riding on public roadways. We investigated the detection of motorcycle riding types using 1) the motorcycle's extracted images, 2) the motorcyclist's extracted images, and 3) the motorcyclist's pose landmarks. You Only Look Once (YOLO) was applied to detect a motorcycle, a motorcyclist, and the landmarks of a motorcyclist from images. The findings indicated that the classification derived from YOLO detectable motorcycles surpassed that of the motorcyclists and their pose landmarks. The VGG16 surpassed MobileNet, CNN, and ResNet50 in classifying normal and dangerous riding. YOLO's efficacy in identifying specific pose landmarks at night was insouciant. Detecting dangerous motorcycling based on the motorcyclists' pose landmarks was ineffective at night. Identifying dangerous motorcycling from the detected motorcycles was the most effective. The findings indicated that YOLO attained an accuracy of 71.09% in motorcycle detection from daytime and nighttime images, whereas VGG16 acquired an accuracy of 98.75% in recognizing dangerous motorcycling.

Enhanced Prediction of Jasmine 105 Rice Growth with RC-ELM and Slow-Release Organic Fertilizers

2025-04-10T11:25:10+07:00

This study explores the use of a Residual Compensation Extreme Learning Machine (RC-ELM) to predict the growth of Jasmine 105 rice, specifically in the context of slow-release organic fertilizers (SROFs). The experiment involved four types of fertilizers: Cow Manure, Filter Cake, Aerated Compost, and a standard chemical control (27-12-6). The macronutrient content of each fertilizer was used as key input variables in the RC-ELM model, with real-time field sensor data providing insights. After extensive preprocessing through normalization and feature engineering, RC-ELM demonstrated superior performance compared to traditional models, such as Linear Regression, Support Vector Machines (SVM), and standard ELM variants. In particular, RC-ELM achieved an R2 = 0.9609, Y=14.982x 103.58 for Aerated Compost, reducing the Mean Squared Error (MSE) by 30%. The results indicate that while organic fertilizers like Aerated Compost may incur higher costs, they offer long-term sustainability benefits, including improved soil fertility. The study further highlights the importance of adopting organic agricultural practices, which align with internationally recognized standards, such as Organic Thailand and IFOAM, for food safety and environmental preservation. These findings underscore the potential of RC-ELM in enhancing crop yield predictions while supporting sustainable farming practices.